English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Apple's AMP1: 104 out of 108 Real-Robot Trials Succeeded, at 280 Seconds per Success

Forum topic · QianXun · 2026-11-11

Summary

A Chinese forum post offers a detailed fact-checked breakdown of Apple's robotics paper "On Agentic Manipulation Policies" (arXiv:2611.09892, v1 submitted 2026-11-06 18:59 UTC, 13 authors including Vladlen Koltun). AMP1 is a reference harness that lets a general-purpose foundation model write, run, and revise its own manipulation code across single-arm, bimanual, and mobile robots, with no grasping primitives, no demonstrations, and no weight updates. On real hardware (12 tasks, 3 Franka Research 3 arms, 9 trials each), AMP1 completed 104 of 108 trials (96.3%), versus 49 for FLUX 3 Action (45.4%) and 30 for π0.5 (27.8%) — a failure-rate reduction of roughly 15x and 20x. The cost: each success averaged 280 seconds wall-clock, 211 of them model inference, versus 62 s (π0.5) and 44 s (FLUX). The post also verifies simulation results (RoboLab-120: 85.6%/82.8% for Gemini 4 Argon/Opus 5.5), emergent behaviors (66% measure containers before placement, 39% diagnose after failure), sensor ablations (82.8% down to 36.8% text-only), persistent-memory learning on RoboDojo (26.4% to 57.6%), and notes the paper's stated limitations: slow inference, no mid-trajectory replanning, and distance from mission-critical deployment.

This is a structured English summary of a detailed Chinese-language fact-check and analysis post about Apple's robotics paper *On Agentic Manipulation Policies* (AMP1). All figures come from the source post and the paper it cites.

Key points

  • Paper identity: arXiv:2611.09892, v1 submitted 2026-11-06 18:59:34 UTC, 13 authors (Apple; including Chen Chen, Omid Bagherieh, Matthias Müller, Vladlen Koltun), categories cs.RO / cs.AI / cs.CV / cs.LG. Code and videos at apple-amp-research.github.io (linked as apple-aiml-research.github.io/ml-amp/ in the post). Note: AMP (plural) names the class of methods; AMP1 is the specific reference harness, not an acronym expansion.
  • Real-robot results: 12 tasks × 3 robot instances × 9 trials = 108 trials. AMP1 succeeded 104/108 (96.3%); FLUX 3 Action: 49 (45.4%); π0.5: 30 (27.8%). Failure rates: 3.7% vs 54.6% and 72.2% — roughly 15× and 20× lower, matching the paper's abstract claim.
  • The cost of reliability: one success averaged 280 seconds (69 s execution + 211 s inference), vs π0.5 at 62 s and FLUX at 44 s. The 280 s mean counts successful trials only. With error recovery disabled, AMP1 dropped to 94/108 (87.0%).
  • How it works: the harness gives a general model three cameras (two shoulder, one wrist), RGB-D, calibration, joint/gripper state, and a stateful Python interpreter. There are no grasp or push primitives — even placement must be self-written. Over 1,200 RoboLab episodes: 66% measured containers before placement, 61% aligned grasp to object axes, 47% reasoned about minor/major axes, 40% read rejection reasons, 39% diagnosed failures before retrying (all gated by a recurrence threshold: ≥5 episodes and ≥2 tasks; behavior stats specific to Opus 5.5). 13 models tested (7 closed + 6 open) — a different "13" from the 13 authors.
  • Notable ablation: adding helper skills *hurt* the four strongest models (up to −7.5 points) but helped the six weakest (up to +17.6 points). The authors' explanation: helper routines hide failures via silent internal retries, so the model can't see atomic-action outcomes and cannot correct in time.
  • Sensor ablation chain (Opus 5.5, RoboLab-120 default): 82.8 (RGB+C+D+T) → 81.9 (no perception tools) → 81.1 (RGB+C) → 75.8 (RGB) → 36.8 (text-only). The proper baseline for 36.8 is the 75.8 RGB tier. Model calls rose as inputs shrank: 34.8 → 38.7 → 46.3 → 53.5.
  • Simulation vs real: RoboLab-120 default-condition success: Gemini 4 Argon 85.6%, Opus 5.5 82.8%; VLA 49.2%, WAM 42.9%. Simulation reports *simulated time* (20–300 s cap) with the simulator paused during inference; real-robot numbers are wall-clock — the two are not directly comparable.
  • CaP-Gym: LIBERO-PRO 95.2%, robosuite 700/700, BEHAVIOR-1K 48/50 (96.0%). HarnessPAI scored 96.5% on LIBERO-PRO but is not zero-shot-equivalent (evolved programs on eval seeds; uses a VLA fine-tuned on the benchmark's verifier).
  • Remaining failures: of Gemini 4 Argon's 173 RoboLab-120 failures, 121 attributable to benchmark artifacts (77 ungraspable geometries/masses, 44 verifier-rejected plausible final states); on the cleaned 95 tasks: 96.2% (914/950).
  • Persistent memory (RoboDojo, 42 bimanual tasks): zero-shot 26.4%; after 5 practice episodes per task within a 48-episode budget, 84.1% (few-shot, failure rate down ~4.6×). More strikingly, after a random pass over all 42 tasks, zero-shot success rose from 26.4% to 57.6% (M42), transferring to unseen tasks without erasing gains; the S4 subset went from 24% to 47%. No weight updates, no expert demonstrations — just a persistent memory folder of notes, code, and data. By the second round, 39/42 tasks had stored code libraries, and model calls roughly halved.
  • Instruction robustness: under ambiguous instructions, Gemini 4 Argon 84.0%, Opus 5.5 79.8%; FLUX 3 Action fell from 42.9% to 28.4%, π0.5 from 28.0% to 15.2%.
  • Stated limitations (from the paper): slow (most of each episode spent waiting on the model); no mid-trajectory replanning; near-quasi-static evaluation scenarios; not mission-critical-ready. The paper's own outlook: "Specialist manipulation policies, such as VLAs and WAMs, are comparably fast but unreliable. AMPs are comparably reliable but slow."
  • Corrections the post makes vs secondhand coverage: the submission date is Nov 6 (not Nov 9, which is an indexing/aggregation date); the paper's "finding hidden objects" means removing occluders, not searching cabinets; the post's kitchen metaphors are the author's own. Not verified: page/figure counts, formal publication record, independent third-party replication.

The takeaway

Reliability is expensive. AMP1 proves that general models can operate robot arms with zero demonstrations, zero primitives, and frozen weights — and know when they erred. It does not yet prove the approach can survive production line timing. Reading 96.3% and 280 seconds on the same line is more honest than quoting either number alone.

Tags

#embodied-ai#robot-manipulation#amp1#apple-research#zero-shot#vla#llm-agents#arxiv-2611-09892

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178637039